Talanta ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Contents lists available at ScienceDirect

Talanta
journal homepage: www.elsevier.com/locate/talanta

Analysis of milk by FT-Raman spectroscopy

Sylwester Mazurek a,n, Roman Szostak a,n, Tomasz Czaja a, Andrzej Zachwieja b
a
Department of Chemistry, University of Wrocław, 14F. Joliot-Curie, 50-383 Wrocław, Poland
b
Faculty of Biology and Animal Sciences, Wrocław University of Environmental and Life Sciences, 38C Chełmońskiego, 50-630 Wrocław, Poland

art ic l e i nf o a b s t r a c t

Article history: Fat, protein, carbohydrates and dry matter were quantified in commercial bovine milk samples, with the
Received 29 October 2014 relative standard errors of prediction (RSEP) in the 3.4–6.1% range, using the partial least squares (PLS)
Received in revised form method based on Raman spectra of liquid milk samples. Results of a better quality were obtained from a
9 March 2015
PLS model derived from IR spectra registered using single reflection ATR diamond accessory, which
Accepted 15 March 2015
yielded RSEP values of 2.4–4.4%. The data indicated IR single reflection ATR spectroscopy and Raman
spectroscopy in combination with multivariate modelling using the PLS method, allowed for the reliable,
Keywords: simultaneous quantitative determination of macronutrients in milk. The low signal to noise ratio of
Milk analysis Raman spectra affects the quality of fat quantification especially for strongly defatted milk samples.
Raman spectroscopy
& 2015 Elsevier B.V. All rights reserved.
ATR
Quantitative analysis
Multivariate calibration

1. Introduction Another vibrational spectroscopy technique, Raman spectroscopy,

is suitable for the analysis of solid and liquid samples, providing
Milk is one of the most important dietary products which con- qualitative and quantitative data [10–12]. It is a non-destructive
tains nearly all the nutrients necessary to sustain life [1]. Being a analytical method, where studied objects require no special pre-
complex natural fluid in which water is a major constituent, milk paration and the spectra are recorded for samples in their native
contains varying quantities of lipids, proteins and carbohydrates as state, significantly simplifying analyses. An important advantage to
well as smaller amounts of minerals and other fat-soluble or water- this method is that Raman data can be collected for substances
soluble components [2]. placed in glass and polymer packaging [13], and may have potential
A number of analytical methods have been developed to deter-
applications in the milk industry.
mine the chemical composition of milk, providing a detailed analysis
Raman spectroscopy has been used for quality control and
of fat, protein, carbohydrates and dry matter which is required for
quantitative analysis of powdered milk constituents [14,15], and to
quality control purposes and consumer information. The traditional
screen samples adulterated with whey [16,17] and melamine
methods used for milk, i.e. Röse-Gottlieb for fat determination,
[18,19]. In the case of liquid milk samples, Raman spectroscopy in
Kjeldahl for proteins and polarimetry or gravimetry for lactose
combination with gel filtration was utilised to detect whey pro-
quantification, have been replaced by modern techniques that could
teins in milk [20]. However, the widespread application of Raman
significantly reduce cost and time needed for milk analysis [3,4].
analysis to liquid milk samples is hampered by the low intensity of
Considering the global production scale of milk and milk-derived
products, simplification of analytical procedures would be extremely its spectra, which are dominated by water bands. The spectral
advantageous. Fifty years ago, infrared (IR) spectroscopy was found contributions of the remaining constituents in milk, accounting for
to be a valuable technique allowing for fast and accurate determi- little more than a few percent of the liquid mass, are weak.
nation of the chemical composition of milk and several IR techniques Raman spectra of liquid milk samples are usually characterised
utilising both mid- (MIR) and near infrared (NIR) spectral ranges by a poor signal to noise (S/N) ratio which makes traditional uni-
were used [5–9]. Infrared spectroscopy was adopted for use in variate analyses difficult to perform. To overcome these limitations
commercially-available milk analysers and has become a standard and extract relevant information from spectral data, chemometric
tool for the quantification of milk components [3,5]. methods can be utilised [21]. The use of multivariate data analysis
techniques often yield robust calibration models, even for noisy
n
systems poorly modelled by univariate techniques [22,23]. Raman
Corresponding authors. Tel.: þ 48 71 3757 238; fax: þ 48 71 3757 420.
E-mail addresses: sylwester.mazurek@chem.uni.wroc.pl (S. Mazurek), spectroscopy in conjunction with PLS modelling was applied to
roman.szostak@chem.uni.wroc.pl (R. Szostak). quantify fat in liquid milk samples [24].

http://dx.doi.org/10.1016/j.talanta.2015.03.024
0039-9140/& 2015 Elsevier B.V. All rights reserved.

Please cite this article as: S. Mazurek, et al., Talanta (2015), http://dx.doi.org/10.1016/j.talanta.2015.03.024i
2 S. Mazurek et al. / Talanta ∎ (∎∎∎∎) ∎∎∎–∎∎∎

This study used FT-Raman spectroscopy to simultaneously models and to perform the quantitative analysis of macronutrients
quantify fat, protein, carbohydrates and dry matter, the main and dry matter in commercial cow milk samples. Principal compo-
components of milk. These results were compared to data nent analysis (PCA) was performed using PLS Toolbox (ver. 6.2,
obtained using a single reflection ATR accessory. Eigenvector Research, Wenatchee, WA, USA) in Matlab (ver. 7.0.4,
MathWorks, Natwick, MA, USA) environment. Fast Fourier transform
(FFT) filter was applied to de-noise Raman spectra recorded with the
2. Materials and methods resolution of 8 cm  1 (an example of the original and smoothed
spectrum is shown in Fig. S1 of supplementary material). Spectral
2.1. Sample preparation data were mean-centred and normalised using the MVN algorithm
[26]. To characterise and compare the predictive abilities of calibra-
Samples (75) were prepared by mixing various commercially- tion models, the relative standard errors of prediction (RSEP) were
available milk, including evaporated, containing 0%, 0.5%, 1.5%, 2.0%, calculated according to the following formula:
3.2% and 7.5% of fat. When necessary, small quantities of water were
also added. Concentrations of 0.5–3.9%, 2.2–4.9% and 3.4–7.0% were ∑in=1 (Ci−CiA )2
RSEP(%) = × 100
obtained for fat, protein and carbohydrates, respectively (a detailed ∑in=1 CiA
2
(1)
composition of the samples, determined using the reference IR
method, is presented in Table S1 of supplementary material). Each A
where C is the component concentration determined by the
sample was divided into 3 portions and quantified by three inde- reference method, C is the concentration found from the model
pendent analytical methods i.e. Raman, ATR and reference IR. Spec- and n is the number of samples. The RSEPcal and RSEPval errors
troscopic measurements were performed without any physical or were calculated for the calibration and validation data sets,
chemical pretreatment of samples, except for maintaining the sam- respectively. The RSEPtest error for each chemical component was
ples at 40 °C. computed for the commercial milk samples. Cross-validation using
the leave-4-out technique was performed to estimate the perfor-
2.2. FT-Raman measurements mance of the models, and the root-mean-square error of cross-
validation (RMSECV) was calculated to select an optimal number
A Nicolet Magna 860 FT-IR spectrometer (Thermo Nicolet, Madi- of factors for the PLS models [21].
son, WI, USA) interfaced with a FT-Raman accessory equipped with
CaF2 beamsplitter and indium–gallium–arsenide (InGaAs) detector
was used to carry out the measurements. Milk samples, placed in a 3. Results and discussion
typical NMR glass tube, were illuminated by a 1.064 mm Nd:YVO4
laser with the power of 400 mW at the sample, and backscattered The chemical composition of milk can vary considerably depend-
radiation was collected. The interferograms were averaged over 512 ing on several factors including species, breed, individuals, stage of
scans, Happ-Genzel apodized and Fourier transformed using a zero- lactation, age, health status and feeding regimen [1]. Raw bovine milk
filling factor of 2 to yield spectra in the 100–3700 cm  1 range at a usually contains 3–5% of fat, 3.5–4% protein, 4.5–5% lactose and  12%
resolution of 8 and 16 cm  1. The interferograms were averaged over of dry matter, while commercial milk samples were characterised by
768 scans for spectra recorded with a resolution of 32 cm  1. Under a lower level of fat due to industrial recovery of this component.
such conditions, each spectrum required approximately 6–10 min to Twenty one local milk samples containing 1.5–3.2% of fat and typical
complete. amounts of protein and carbohydrates were selected for analysis.

2.3. FTIR ATR measurements

3.1. FT-Raman analysis
ATR spectra were recorded using a single reflection Golden Gate
Representative Raman milk spectra are shown in Fig. 1. These
(Specac, Slough, UK) diamond accessory mounted in the Nicolet iS50
spectra were dominated by water scattering, while the remaining
FTIR spectrometer (Thermo Fisher Scientific, Madison, WI, USA). A
milk constituents were responsible for a number of low intensity
KBr beamsplitter and DTGS detector were used for infrared mea-
surements. The spectra of milk samples in the 400–4000 cm  1 range
with a resolution of 4 cm  1 were obtained (128 scans). Similar to
Raman experiments, Happ-Genzel apodization and Fourier trans-
6
formation of interferograms using a zero filling factor of 2 were
applied to obtain IR spectra. Under such conditions, it took
approximately 2–3 min to register the spectrum. ATR measurements
were performed in a similar manner to those described by Iñón et al.
Raman intensity

4 f
[4], though a single reflection diamond accessory was used instead of
a 6-reflection ZnSe accessory.
e

2.4. Reference analysis d

2
c
Comparative analyses of the chemical composition of milk were
b
performed with the B-150 IR milk analyser (Bentley Instruments,
Chaska, MI, USA). Fat, protein, carbohydrates and dry matter content a
was determined according to the ISO 9622 procedure [25]. 0

3500 3000 2500 2000 1500 1000 500

2.5. Software and numerical data treatment -1
Wavenumber [cm ]
Nicolet TQ Analyst Version 7 (Thermo Fisher Scientific, Madison, Fig. 1. FT Raman spectra of commercial milks containing (a) 0%, (b) 0.5%, (c) 1.5%,
WI, USA) chemometric software was used to construct the PLS (d) 2.0%, (e) 3.2% and (f) 7.5% (evaporated) of fat; spectra are offset for clarity.

Please cite this article as: S. Mazurek, et al., Talanta (2015), http://dx.doi.org/10.1016/j.talanta.2015.03.024i
 S. Mazurek et al. / Talanta ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3

bands present in different regions of the spectra. The S/N values Table 1
obtained for milk Raman spectra were at least 10-fold lower than Calibration parameters for milk component determination.

those found for spectrum of pure lactose (FT-Raman spectra of Component Parameter FT Raman FTIR ATR
water, lactose, fat and casein are shown in Fig. S2 of supplemen-
tary material). 8 cm  1 16 cm  1 32 cm  1

All PLS calibration models were constructed using the Raman

Fat R 0.994 0.993 0.989 0.995
spectra of the 75 laboratory prepared milk samples. Sixty samples Rcv 0.964 0.969 0.968 0.977
were used for training purposes while 15 constituted an external RSEPcal (%) 5.09 4.66 6.27 3.28
RSEPval (%) 6.41 5.91 7.05 3.47
validation set (see Table S1 of supplementary material), and the
selection of validation samples was supported by the PCA.
Principal component analysis was used to assess spectral uni- Protein R 0.985 0.980 0.930 0.991
Rcv 0.849 0.873 0.771 0.988
formity between calibration samples and commercial samples, RSEPcal (%) 3.37 3.43 6.32 1.86
and Fig. 2 shows the score plots obtained as a result of PCA RSEPval (%) 4.47 3.80 6.79 2.28
decomposition of the Raman intensity matrix. The positions of real
milk samples in the variability space of the few first principal Carbohydrates R 0.983 0.980 0.938 0.993
components confirmed that spectra of the calibration mixtures Rcv 0.811 0.889 0.803 0.990
RSEPcal (%) 3.19 2.82 5.59 1.86
reflect the spectral variability characterising commercial milk
RSEPval (%) 4.35 3.10 5.79 1.93
samples. These data also indicated that the laboratory mixtures
reflect the chemical structure of real milk samples. A preliminary
Dry matter R 0.976 0.994 0.987 0.995
analysis performed for samples prepared by dissolving lactose, Rcv 0.897 0.971 0.957 0.990
casein and selected other milk constituents in water resulted in RSEPcal (%) 4.62 2.13 3.24 2.36
RSEPval (%) 5.19 2.65 3.42 2.47
milk-like samples poorly mimicking liquid milk. These samples
were outliers in the PCA score plots obtained from spectral data.

Spectral ranges of 370–860, 1020–1809 and 2800–2953 cm  1,

including the most characteristic milk bands were selected for
2 modelling purposes. The number of latent PLS variables were
determined on the basis of RMSECV plots (Fig. S3 of supplementary
material) and were set to 3, except for protein and carbohydrates in
the case of spectra recorded with the resolution of 8 cm  1, when
Scores on PC 2 (4.57%)

4 PLS factors were used.

The detailed PLS modelling results are summarised in Table 1.
0 Although a difference in the quality of models based on Raman
spectra registered with different resolution is visible, it is not pro-
nounced. Correlation coefficient for the obtained calibration curves
-1 varied in the range 0.930-0.994. The RSEP errors were generally
lower for models based on Raman spectra registered with the reso-
lution of 16 cm  1. These values varied in the 2.1–4.7% and 2.6–5.9%
-2 range for calibration and validation samples, respectively (Table 1).
Raman spectra measured with the resolution of 32 cm  1 are
characterised by a higher S/N ratio in comparison with those
-6 -4 -2 0 2 4 6
registered with a better resolution and they can be recorded faster.
Scores on PC 1 (40.42%)
Unfortunately, apparently too much of the specific spectral infor-
mation is lost and final calibration models, based on these spectra,
give higher quantification errors. On the other side, spectra
2 registered with the resolution of 8 cm  1 are too noisy which
results in less robust calibration models. Calibration curves, based
on Raman spectra recorded with the resolution of 16 cm  1, and
plots of the relative errors obtained for milk components are
Scores on PC 3 (3.55%)

1
presented in Fig. 3.
Twenty one commercial milk samples containing 1.5% (n ¼7),
2.0% (n ¼7) and 3.2% (n ¼7) fat were analysed applying the
0
developed calibration models. The RSEPtest errors in the range of
5.3–5.8% for fat, 5.6–6.1% for protein, 3.5–4.8% for carbohydrates
and 3.4–4.8% for dry matter were found (Fig. 4).
-1 Attempts to determine amount of fat in milk samples containing
less than 1% fat resulted in unacceptable high quantification errors
due to the weakness of respective bands in the Raman spectra.
-2
3.2. Repeatability of quantification
-6 -4 -2 0 2 4 6
Scores on PC 1 (40.42%) In a separate step, a set of 64 samples of commercial milk samples
Fig. 2. PCA analysis: score plots for milk samples with the 95% confidence interval; containing 3.2% of fat were analysed using the developed calibration
squares-calibration samples, circles-validation set, triangles-commercial samples. models (the scores plots of PCA modelling for these samples are

Please cite this article as: S. Mazurek, et al., Talanta (2015), http://dx.doi.org/10.1016/j.talanta.2015.03.024i
4 S. Mazurek et al. / Talanta ∎ (∎∎∎∎) ∎∎∎–∎∎∎

4
fat
R=0.9934

10
3
Calculated C [%]

Difference [%]
2 0

1 -10

1 2 3 4 1 2 3 4
Actual C [%] Actual C [%]

5
protein
R=0.9805

5
4
Calculated C [%]

Difference [%]
0

3
-5

2
2 3 4 5 2 3 4 5
Actual C [%] Actual C [%]

7 carbohydrates
R=0.9803
5
6
Calculated C [%]

Difference [%]

0
5

4 -5

4 5 6 7 4 5 6 7
Actual C [%] Actual C [%]

16
dry matter 6
R=0.9937

3
Calculated C [%]

Difference [%]

12
0

-3
8

-6

8 12 16 8 12 16
Actual C [%] Actual C [%]

Fig. 3. Calibration curves and relative errors for milk components on the basis of Raman spectra registered with the resolution of 16 cm  1; squares-calibration samples,
circles-validation samples.

Please cite this article as: S. Mazurek, et al., Talanta (2015), http://dx.doi.org/10.1016/j.talanta.2015.03.024i
 S. Mazurek et al. / Talanta ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 5

-1 another effective tool for the simultaneous quantitative determi-

Raman 8 cm
Raman 16 cm
-1 nation of macronutrients in milk. Fat, protein, carbohydrates and
6 Raman 32 cm
-1
dry matter were quantified in 21 commercial milk samples with
FTIR ATR
RSEPtest errors of 3.4–5.6% relative to the values found applying the
IR reference method. Results of a better quality were obtained
from a separate PLS model based on IR spectra registered using a
single reflection ATR diamond accessory, yielding RSEPtest values
RSEP [%]

of 2.4–4.4%.
3

Acknowledgments

Authors thank Ms. A. Łoza and Ms. M. Lenarska from the Wrocław
University of Environmental and Life Sciences for performing the
reference analysis of milk samples.
0
fat protein carbohydrates dry matter

Fig. 4. Raman and ATR RSEPtest quantification errors for macronutrients and dry Appendix A. Supplementary material
matter in commercial milk samples (n¼21).
Supplementary data associated with this article can be found
shown in Fig S4 of supplementary material). Milk components were in the online version at http://dx.doi.org/10.1016/j.talanta.2015.03.
quantified with the mean relative error values in the range 4.2–5.8% 024.
for fat, 3.8–4.6% for protein, 3.9–4.8% for carbohydrates and 2.4–2.8%
for dry matter (Fig. S5 of supplementary material).
References
3.3. FTIR ATR analysis
[1] R.G. Jensen, Handbook of Milk Composition, Academic Press, London, 1995.
Simultaneous with Raman measurements, single reflection ATR [2] A. Varnam, J.P. Sutherland, Milk and Milk Products: Technology, Chemistry and
Microbiology, Chapman & Hall, New York, 1994.
spectra of the same milk samples were recorded (two sets of milk [3] D.A. Biggs, J. Dairy Sci. 50 (1967) 799–803.
IR spectra, raw and after water spectrum subtraction, are shown in [4] F.A. Iñón, S. Garrigues, M. de la Guardia, Anal. Chim. Acta 513 (2004) 401–412.
Figs. S6 and S7 of supplementary material, respectively). Applying [5] J.D.S. Goulden, J. Dairy Sci. 31 (1964) 273–284.
[6] R. Tsenkova, S. Atanassova, K. Itoh, Y. Ozaki, K. Toyoda, J. Anim. Sci. 78 (2000)
methodology analogous to that used during Raman data analysis, 515–522.
PLS modelling was performed (calibration curves and relative [7] H. Soyeurt, P. Dardenne, F. Dehareng, G. Lognay, D. Veselko, M. Marlier,
errors plots for milk constituents were shown in Fig. S8 of sup- C. Bertozzi, P. Mayeres, N. Gengler, J. Dairy Sci. 89 (2006) 3690–3695.
[8] B. Aernouts, E. Polshin, J. Lammertyn, W. Saeys, J. Dairy Sci. 94 (2011)
plementary material). The following spectral ranges were utilised: 5315–5329.
800–1017, 1063–1634 and 1732–1760 cm  1. Correlation coefficient [9] A. Melfsen, E. Hartung, A. Haeussermann, J. Dairy Sci. 95 (2012) 6465–6476.
values determined on the basis of ATR data varied from 0.991 to [10] E.C.Y. Li-Chan, Trends Food Sci. Technol. 7 (1996) 361–370.
[11] R. Szostak, S. Mazurek, Analyst 127 (2002) 144–148.
0.995, while RSEP errors were 1.9–3.3% and 1.9–3.5% for calibra-
[12] L.G. Thygesen, M.M. Løkke, E. Micklander, S.B. Engelsen, Trends Food Sci.
tion and validation samples, respectively. Representative results Technol. 14 (2003) 50–57.
for the ATR PLS modelling are shown in Table 1. Quantitative [13] A.M. Tudora, C.D. Melia, J.S. Binns, P.J. Hendra, S. Church, M.C. Davies, J. Pharm.
analyses of commercial samples resulted in the RSEPtest error Biomed. Anal. 8 (1990) 717–720.
[14] J. Moros, S. Garrigues, M. de la Guardia, Anal. Chim. Acta 593 (2007) 30–38.
values of 2.4–4.4% (Fig. 4). [15] C.M. McGoverin, A.S.S. Clark, S.E. Holroyd, K.C. Gordon, Anal. Chim. Acta 673
(2010) 26–32.
3.4. Comparison of the techniques [16] M.R. Almeida, K. Oliveira, R. Stephani, L.F.C. de Oliveira, J. Raman Spectrosc. 42
(2011) 1548–1552.
[17] J. Qin, K. Chao, M.S. Kim, Food Chem. 138 (2013) 998–1007.
The relatively low intensity of Raman signals given by the milk [18] S. Okazaki, M. Hiramatsu, K. Gonmori, O. Suzuki, A.T. Tu, Forensic Toxicol. 27
samples suggested that longer collection times, in comparison (2009) 94–97.
[19] Y. Cheng, Y. Dong, J. Wu, X. Yang, H. Bai, H. Zheng, D. Ren, Y. Zou, M. Li, J. Food
with the IR technique, were required to obtain acceptable S/N Compos. Anal. 23 (2010) 199–202.
ratios. Having in mind the quality differences between the Raman [20] M. Liang, V.Y.T. Chen, H.L. Chen, W. Chen, Talanta 69 (2006) 1269–1277.
and ATR spectra (see Fig. 1 and Figs. S6 and S9 of supplementary [21] B.M. Wise, B.R. Kowalski, in: F. McLennan, B.R. Kowalski (Eds.), Process in
Analytical Chemistry, Chapman & Hall, London, 1995, pp. 259–312.
material) it is not surprising that the prediction errors for carbo- [22] R. Kramer, Chemometric Techniques for Quantitative Analysis, Marcel Dekker,
hydrates quantification based on the Raman spectra appeared to New York, 1998.
be 50% and for protein determination even two times higher when [23] I. Stefanov, V. Baeten, B. De Baets, V. Fievez, Talanta 112 (2013) 101–110.
[24] R.M. El-Abassy, P.J. Eeravuchira, P. Donfack, B. von der Kammer, A. Materny,
compared with those obtained from ATR data (Fig. 4). The obvious
Appl. Spectrosc. 66 (2012) 538–544.
advantage of Raman spectroscopy over IR technique is that it can [25] PN-ISO 9622:2006 (IDF 141), Milk and Liquid Milk Products—Guidelines for
be applied to the on-line monitoring of the chemical composition the Application of Mid-infrared Spectrometry, Polish Committee for Standar-
of milk at different stages of the industrial milk processing. dization, Warsaw, 2006.
[26] R. Szostak, S. Mazurek, Vib. Spectrosc. 49 (2009) 298–302.

4. Conclusions

The results indicated that multivariate modelling using the PLS

method based on Raman spectra of liquid milk samples was

Please cite this article as: S. Mazurek, et al., Talanta (2015), http://dx.doi.org/10.1016/j.talanta.2015.03.024i